Enhanced Neural Distinguisher Model for Efficient Differential Cryptanalysis

At CRYPTO 2019, Gohr applied deep learning to differential cryptanalysis of SPECK32/64, achieving identification accuracy surpassing that of traditional differential distinguishers. This achievement offers new perspectives for data security and privacy protection in the Internet of Things (IoT). However, existing research still faces challenges, such as limited model accuracy and excessive computational resource consumption. To address these issues, we propose a novel enhanced model of differential neural distinguishers that balances high accuracy with low computational overhead. Initially, an innovative data feature extraction strategy is designed by introducing the skip connection mechanism to effectively integrate both linear and nonlinear features extracted from the raw data. This allows the model to better approximate the internal mechanisms of cryptographic algorithms. Subsequently, based on the positional relationships of nonlinear components within round functions and the diffusion properties of linear components, an original input data format selection strategy is proposed. We employ the multipair data augmentation strategy, significantly improving the model’s identification accuracy and generalization capabilities. Additionally, we pioneer the integration of an efficient channel attention (ECA) module, to curtail the number of residual blocks required, thereby effectively reducing computational load. Furthermore, leveraging the algebraic expressions of cryptographic ciphers and the propagation characteristics of differential features, we develop a fast neutral bit search algorithm that enhances the efficiency of the key recovery process. Taking SIMON32/64 as an example, we successfully demonstrate a key recovery attack for 16 rounds with an accuracy rate of 80%.